Re: [PATCH] asm-generic: io: Fix ioport_map() for !CONFIG_GENERIC_IOMAP && CONFIG_INDIRECT_PIO

[Andrew Murray]

On Mon, Sep 17, 2018 at 03:42:32PM +0100, John Garry wrote:
> - dead e-mail addresses (Zhichang, Gabriele)
> 
> On 13/09/2018 13:48, Andrew Murray wrote:
> 
> Hi Andrew,
> 
> > The !CONFIG_GENERIC_IOMAP version of ioport_map uses MMIO_UPPER_LIMIT to
> > prevent users from making I/O accesses outside the expected I/O range -
> > however it erroneously treats MMIO_UPPER_LIMIT as a mask which is
> > contradictory to its other users.
> > 
> > The introduction of CONFIG_INDIRECT_PIO, which subtracts an arbitrary
> > amount from IO_SPACE_LIMIT to form MMIO_UPPER_LIMIT, results in ioport_map
> > mangling the given port rather than capping it.
> > 
> > We address this by aligning more closely with the CONFIG_GENERIC_IOMAP
> > implementation of ioport_map by using the comparison operator and
> > returning NULL where the port exceeds MMIO_UPPER_LIMIT. Though note that
> > we preserve the existing behavior of masking with IO_SPACE_LIMIT such that
> > we don't break existing buggy drivers that somehow rely on this masking.
> 
> I wouldn't say any drivers rely on this - for the only device driver which
> uses the "Indirect" IO space region above MMIO_UPPER_LIMIT (HiSilicon LPC),
> no child device driver for that host uses ioport_map() [those being ipmi si
> and 8250 generic+of drivers].

I was really referring to the existing !CONFIG_GENERIC_IOMAP && 
!CONFIG_INDIRECT_PIO use cases where there may be drivers (however unlikely)
that provide ioport_map an incorrect address which, due to the masking, gets
converted into a valid address. Returning NULL for these would result in new
run-time errors therefore it seemed safer to change this to support the new
"indirect IO" whilst not breaking existing bad drivers.

A more correct implementation would always return NULL if
port > IO_SPACE_LIMIT - it would fully align it with the CONFIG_GENERIC_IOMAP
implementation - in my view these two implementations should behave the same
with respect to error handling - at the moment they don't.


> 
> Regardless of that, it seems better to return NULL when the port is
> out-of-range, rather than masking it.
> 
> Cheers
> 
> > 
> > Fixes: 5745392e0c2b ("PCI: Apply the new generic I/O management on PCI IO hosts")
> > Reported-by: Will Deacon <will.deacon@xxxxxxx>
> > Signed-off-by: Andrew Murray <andrew.murray@xxxxxxx>
> 
> Reviewed-by: John Garry <john.garry@xxxxxxxxxx>

Thanks for the review.

Andrew Murray

> 
> > ---
> >  include/asm-generic/io.h | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
> > index 66d1d45..d356f80 100644
> > --- a/include/asm-generic/io.h
> > +++ b/include/asm-generic/io.h
> > @@ -1026,7 +1026,8 @@ static inline void __iomem *ioremap_wt(phys_addr_t offset, size_t size)
> >  #define ioport_map ioport_map
> >  static inline void __iomem *ioport_map(unsigned long port, unsigned int nr)
> >  {
> > -	return PCI_IOBASE + (port & MMIO_UPPER_LIMIT);
> > +	port &= IO_SPACE_LIMIT;
> > +	return (port > MMIO_UPPER_LIMIT) ? NULL : PCI_IOBASE + port;
> >  }
> >  #endif
> > 
> > 
> 
>